Hybrid Search Infrastructure
Running hybrid search in production means combining a vector index and a keyword index, then merging their results into one ranking.
Prerequisites
Overview
Hybrid search needs two indexes running in parallel — a vector index for semantic similarity and a keyword (typically BM25-style) index for exact term matches — plus a merging strategy that combines their two different scoring systems into one ranking.
Where It Fits
Query
Vector + Keyword Indexes
Both queried in parallel.
Score Merging
Ranked Results
Key Points
- Two indexes
- Hybrid search typically runs a vector index and a keyword index side by side rather than one system doing both natively.
- Score fusion
- Vector similarity and keyword relevance scores are on different scales — a fusion method like reciprocal rank fusion combines them into one ranking.
- Platform support
- Some vector databases (Weaviate, OpenSearch) support hybrid search natively; others need it built manually across two systems.
Interview Question
Why can’t you just average a vector similarity score and a keyword relevance score to combine them?
The two scores come from unrelated scales and distributions — a raw average would be dominated by whichever happens to have a larger numeric range. A rank-based fusion method, like reciprocal rank fusion, combines the two result orderings rather than their raw scores, which is more robust to that mismatch.
Explain It in 30 Seconds
Hybrid search in production runs a vector index and a keyword index in parallel and merges their two differently-scaled result sets — usually with a rank-based fusion method rather than a raw score average.
Real-World Stack
Technologies commonly used to implement this in production.